AI Search Visibility (GEO)

Make a site citable by AI engines — crawler allowlist, @graph @id, IndexNow, answer-first pages. Use when a site is invisible to ChatGPT. Trigger on "GEO audit", "llms.txt".

Being crawled and being cited are two different pipelines. Most sites lose the second one while trying to configure the first.

Access, data, network

Item	Default
Network calls	Read-only GETs to public URLs you name — your own robots.txt/sitemap.xml (§1, §8), public competitor pages (§9), a third-party SEO index (§10 source 1). Engine probes (§10 source 4) are run by hand or via the vendor's official API — never a scripted consumer UI. Plus one opt-in POST of your own public sitemap URLs via scripts/indexnow-ping.js. Nothing else. No authenticated call, no page content or analytics uploaded, nothing written to the network except that one POST.
Credentials	None. INDEXNOW_KEY is read from env, never written into the repo.
Data persisted	Nothing. Audit output is a file you write, where you choose.
Third-party accounts	Search consoles, knowledge base, business listing, directories — manual, human-operated. This skill never logs into them.
Personal data	Public pages only. Never reads, inventories, scrapes or republishes visitor, customer or prospect records. Internal demand signals enter only as an aggregated theme list produced by a human (§10 source 3) — this skill never opens a customer-data store. Engine ToS and GDPR (or local equivalent) bind you before anything here does.
Authorship	Any copy this skill drafts is a draft. The named human author reviews and endorses before publish (§11). Never a byline on unreviewed text, never an invented person.

When to Use

Trigger	Action
"we're not cited by ChatGPT / Perplexity"	Instrument first, in order: §1 taxonomy → §3 @graph → §4 sitemap → §5 IndexNow
"should we block AI crawlers?"	§1. Never one wholesale rule
"GEO audit" / "why do we rank nowhere"	Audit: §7 inventory → §8 signals → §9 benchmark → §10 scoring /100
"rewrite this page" / "it isn't answer-first"	§11 template + §12 forbidden list
"this article is out of date"	references/playbooks.md §14
"our two sites compete with each other"	references/playbooks.md §15

Placeholders

Placeholder	Example	Meaning
${SITE}	acme-corp.com	Product brand — short pages, urgency intent
${SITE2}	acme-experts.com	Expertise brand — long canonical pillars
${INDEXNOW_KEY}	env var, 32-hex	Env only. Never hardcode, never commit, never paste into docs
#organization, #founder	https://${SITE}/#founder	Stable @id fragments. Never a person's name in the fragment

1. Crawler taxonomy — the thing everyone gets wrong

Three distinct classes of user-agent, from three different pipelines. A blanket "block AI bots" rule, or a WAF pattern matching *bot*, kills classes 2 and 3 in order to opt out of class 1 — and it is invisible to whoever shipped it.

Class	What it does	Blocking it costs you	Named agents
Training crawler	Bulk-collects corpus for model training	Nothing today. Possibly long-term recall of your brand	GPTBot, ClaudeBot, Meta-ExternalAgent, CCBot
Search crawler	Builds the answer index the assistant retrieves from	Citability — you vanish from answers entirely	OAI-SearchBot, Claude-SearchBot, PerplexityBot, Applebot
On-demand fetcher	Fetches a URL because a user asked the assistant to read it	The user pastes your link and gets "I can't access that page"	ChatGPT-User, Claude-User, Perplexity-User
Not crawlers	Google-Extended / Applebot-Extended are training opt-out controls	Disallowing them does not remove you from that vendor's search index; allowing them does not add you	—

Agent names are public facts, but re-verify each against the vendor's current docs before shipping — this list moves every quarter. Full file: templates/robots.txt.

Groups do not inherit, and a group runs until the next User-agent: line (RFC 9309). An agent matching a named group ignores User-agent: * entirely, and blank lines do not end a group — so an internal Disallow written under the wrong heading silently belongs to the agent above it, and applies to nobody else. Repeat every internal Disallow in every group. And put Disallow before Allow: / in each group: RFC 9309 says the longest match wins, so order shouldn't matter — but plenty of live parsers take the first match instead, hit Allow: /, and serve the internal tree anyway. Disallow-first is correct under both readings, and it is free insurance against a crawler you'll never audit.

text

User-agent: OAI-SearchBot     # several agents may share one group
User-agent: Claude-SearchBot
Disallow: /internal/          # BEFORE Allow, and repeated in EVERY group
Allow: /

User-agent: Bytespider        # abusive scrapers: named block only
Disallow: /                   # covers everything; no internal lines needed

Sitemap: https://${SITE}/sitemap.xml

robots.txt is public: a Disallow advertises a path, it does not protect it. Never list a path whose existence is itself sensitive — use auth or a noindex header — and keep host-specific paths out, they name your vendor to every reader.

bash

# Verify what you shipped, not what you meant to. Check the WAF too — robots.txt is not the only gate.
curl -s "https://${SITE}/robots.txt" | grep -c "^User-agent: OAI-SearchBot"
# Output: 1
# Every allowlisted group must carry the internal Disallow: count must equal the number of groups.
curl -s "https://${SITE}/robots.txt" | grep -c "^Disallow: /internal/"
# Output: 5

2. llms.txt — a brief for an engine reading you cold; the structure is the value

Copy the skeleton — six sections, in order, with the rule for each: templates/llms.txt. Two that decide whether it works: core pages must be annotated with what each one answers (a bare link list is worthless), and proof must be third-party and dated (no superlatives, every differentiator falsifiable).

Never include vendor stack, internal pricing logic, personal email, private booking links or unpublished client names. This is a public file, reproduced verbatim by anything that reads it: everything in it must already be published on the site.

3. Schema.org — one @graph, stable @id (the real technical lever)

Stacking independent JSON-LD blocks per page hands the engine three Organization nodes with three names and no way to know they are the same company. One @graph where entities reference each other by {"@id": ...} gives it one entity. The @id is the primary key: reuse the same one on every page — homepage, pricing, and the blog template — and cross-page entity resolution starts working. Full fictional graph: templates/graph.json (Organization · Person · WebSite · WebPage · BreadcrumbList · Service · FAQPage, all cross-referenced). The pattern — note the article template references the same two @ids rather than re-inlining them:

json

{ "@type": "Organization", "@id": "https://acme-corp.com/#organization",
  "founder": { "@id": "https://acme-corp.com/#founder" } },
{ "@type": "BlogPosting", "@id": "https://acme-corp.com/blog/post/#post",
  "author":    { "@id": "https://acme-corp.com/#founder" },
  "publisher": { "@id": "https://acme-corp.com/#organization" } }

Rule	Why
Article templates reference author/publisher by @id only — never re-inline name + logo	An inlined author drifts from the homepage's and becomes a second, weaker entity
@id fragments are role-based (#founder), never patronymic	Permanent keys: they must survive a person leaving and carry no PII
sameAs → knowledge-base item + one canonical profile. Not eight social links	sameAs disambiguates; it is not link-building
One @graph per page, not one per component	Contradictory duplicates are worse than no markup at all
Every dateModified matches a visible date in the rendered HTML	A date the engine can't read on the page is a claim it discounts

4. Sitemap — reciprocal hreflang

Every URL declares all its alternates including itself, plus x-default. Non-reciprocal alternates are silently dropped: A→B without B→A groups neither.

xml

<url>
  <loc>https://acme-corp.com/fr/pricing.html</loc>
  <lastmod>2026-02-01</lastmod>
  <xhtml:link rel="alternate" hreflang="fr" href="https://acme-corp.com/fr/pricing.html"/>
  <xhtml:link rel="alternate" hreflang="en" href="https://acme-corp.com/en/pricing.html"/>
  <xhtml:link rel="alternate" hreflang="x-default" href="https://acme-corp.com/fr/pricing.html"/>
</url>

The fr self-reference is required, not redundant. Repeat the identical block on /en/pricing.html. x-default points at the primary market, identically in every block of the group.

5. IndexNow — most direct lever, wired so it can never break a deploy

IndexNow pushes URLs into the index feeding several assistants' search layer; everything else here is pull. Ship scripts/indexnow-ping.js — non-negotiable properties:

Property	Implementation
Key from env	const KEY = process.env.INDEXNOW_KEY. The key is public by design — but hardcoded into script + docs + build passthrough, a rotatable key becomes three commits you have to hunt down
No throw path	req.on('error') → resolve(). HTTP ≥400 → log, resolve(). Top-level .catch() → warn, exit 0
Timeout	setTimeout(10000) → destroy + resolve(). Never hang a build on a third party
Chunking	1000 URLs per POST, well inside the protocol limit
Build command	… && node scripts/indexnow-ping.js || true — the last line of defence
Sitemap source	SITEMAP_PATH, default ./site/sitemap.xml relative to the working directory — not to the skill's install directory. Set it, or pass URLs as arguments
Entity decoding	<loc> is XML: & is escaped as &amp;. Decode before POSTing or you submit a URL that never existed

bash

export INDEXNOW_KEY=$(node -e "console.log(require('crypto').randomBytes(16).toString('hex'))")
export INDEXNOW_HOST=acme-corp.com                 # store both in the host's env, not the repo
export SITEMAP_PATH=./site/sitemap.xml             # or pass URLs as arguments instead
printf '%s' "$INDEXNOW_KEY" > "site/${INDEXNOW_KEY}.txt"   # serve at domain root
node scripts/indexnow-ping.js
# Output: [indexnow] Submitting 24 URLs to api.indexnow.org…
# Output: [indexnow] HTTP 200 (24 URLs)

Never commit the key file. The filename is the key, so committing it commits the key — and it survives rotation in the git history, which is the exact problem env-only handling exists to avoid. Add site/*.txt (or the exact key filename) to .gitignore and emit it from the build step that reads $INDEXNOW_KEY. Public by design is not the same as public forever, in your history, after you rotated it.

The gotcha that costs a day. You will get HTTP 403 on every call while the key file is correctly hosted, correctly named, and byte-matches the payload. Nothing is wrong with your code: the endpoint 403s until the domain is validated on the engine's webmaster console — a key generated locally is not a key declared there. Do playbooks.md §16 step 4, then re-ping. Meanwhile the sitemap covers crawl on the normal schedule, so you are losing latency, not coverage. Do not "fix" the script.

6. External consoles → references/playbooks.md §16

Human-operated, every step; this skill holds no console credentials. Causal order — domain property (TXT DNS) → sitemap → secondary console by import → declare the IndexNow key (clears the §5 403) → knowledge-base item → business listing → directories with byte-identical NAP.

7. Output format — the inventory table

One row per URL, before any opinion is formed. Count words on rendered text, not source. Three verdicts, grouped by hub: thin (<900 words on a pillar intent) · dated (facts older than the last change in the underlying reality) · orphan (no internal inbound link). Close the audit with the verdict line, verbatim:

text

| URL           | H1 (verbatim)              | Words | Verdict                          |
| /hub/topic-a/ | "Everything about topic A" |  3500 | dated — stats 11 yrs old, 0 FAQ  |
| /hub/topic-b/ | "Topic B: what to do"      |   650 | thin — critical                  |
| /standalone/  | "Topic C"                  |  1900 | orphan — 0 inbound internal links|

GEO SCORE: 47/100 — 3 Tier S, 11 refresh, 0 FAQPage sitewide

8. Technical signals — check every one, every time

Signal	Pass condition
JSON-LD by type	Article/BlogPosting, FAQPage, Person, Organization/LocalBusiness, BreadcrumbList — present and cross-referenced by @id. Zero FAQPage across an entire site is the most common finding, and the cheapest to fix. Note: FAQPage stopped producing SERP rich results for most sites in 2023 — ship it for extraction by generative engines (self-contained 40-60-word Q/A), not for SERP real estate. Different pipeline, still worth it
Visible "Updated on"	Rendered under the H1. Not dateModified alone
Canonical	Self-referencing on canonical pages; pointing at the canonical domain for shared subtopics
Open Graph	og:title/description/url/image (1200×630) + twitter:card
Sitemap	Reachable, real lastmod, reciprocal alternates, referenced from robots.txt
Robots	Named allowlist per §1, no wholesale AI block. Internal trees disallowed in every named group (groups don't inherit from *) and before each Allow: /. Count the Disallow lines: one per group, or it doesn't apply where you think
Core Web Vitals	Measured, not assumed — mobile field data

9. Competitive benchmark

Score each competitor's key page on six axes — depth, FAQ schema, named author, visible update date, reviews, price shown — then read your own position on the queries that pay (commercial intent, not vanity head terms). Table shape and a worked example: references/playbooks.md §17. Name competitors only as A/B/C in anything that can leave your machine — a benchmark naming real firms is a document you cannot share with the client it was built for. If none ships FAQPage schema, that is a time-boxed arbitrage window, not a moat — it closes the quarter one of them notices.

10. Scoring /100 — and the 5 sources when you have no Search Console

No console access at kickoff is the normal case, not the exception — that is exactly what the grid is for.

Criterion	Weight
Current organic traffic	25
Query volume potential	20
Age / obsolescence of content (penalty)	15
Alignment with what you actually sell	15
Already cited by an AI engine	15
Technical quality (length, structure, schema)	10

These weights are a starting point to adjust to your business, not a calibrated constant. What matters is that the grid is written down before you score, not that these exact numbers are optimal — an explicit grid you argue with beats an implicit one you don't.

Thresholds: > 60 → top-20 refresh list · > 75 AND untouched > 18 months AND underlying facts changed since publication → Tier S, urgent. Scores derived from the sources below are provisional — re-rank once console data lands.

Cross these 5 sources to reconstruct the real top without console data:

1. Third-party SEO index — pages ranking on >1 000/mo keywords, pages holding backlinks, pages with the highest traffic value.
2. Whatever analytics already exists — sessions, conversions, >2 min dwell. A partial beacon beats guessing.
3. Internal demand signals — themes only. Work from an aggregated, already-anonymised list of recurring questions that a human hands you. Never from raw inboxes, DMs, CRM records or call recordings: this skill does not read a customer-data store. What people actually ask outranks what you assume they search — a theme list carries that signal without the records.
4. AI engine probes — run by hand, or through the vendor's official API. Run the target queries against each assistant, record who gets cited. Never script the consumer UI: it breaches most engines' terms, and a human reading five answers is faster than the scraper you'd maintain. Competitor cited + you absent = a scored gap, not an anecdote.
5. Your own crawl — sitemap lastmod, word count, inbound internal link count. Surfaces thin/dated/orphan mechanically.

11. The answer-first template — imposed order of 13 blocks, no reordering

#	Block	Spec
1	H1 as a question	The question the user types, verbatim
2	Date + author block	"Updated on DD Month YYYY by [NAME], [credential]"
3	Direct answer, 2-3 sentences	Contains a number and names its source. This is the block that gets quoted
4	TL;DR, 4-6 bullets	Consequences, deadlines, procedure, recourse
5	Table	Thresholds / tiers / amounts. Real <table>, never an image
6	H2s, all questions	Each answerable in 40-60 self-contained words
7	Numbered steps	For anything actionable
8	Common mistakes	"What to avoid" — highest-engagement block on most pages
9	FAQ, 8-12 Q/A	<dl><dt><dd> + FAQPage JSON-LD
10	Contextual CTA	Tied to this page's problem. Never a generic download
11	Hub-and-spoke links	≥5 internal, to sibling and child pages
12	Author block	Credentials + years + link to a real bio page
13	JSON-LD	Article + FAQPage + Person + Organization, by @id per §3

Attribution rule. A named author is a claim of human authorship. Any copy this skill drafts is a draft: the named person must read, correct and approve it before publish. Never put a byline on text that person has not endorsed, and never invent a person, a credential or a bio page — block 2 and block 12 are worthless the moment they are fiction, and an engine that catches one fabricated author discounts the whole domain.

Typography rules: formal register throughout · no jargon without an inline definition, re-defined on every page (a page is an entry point, not chapter 4) · numbers always precise ("6 to 36 months", never "several months") · every claim carries its source reference · update date under the H1 and at the foot · keep commercial claims sparse on a service page (3 is already a lot) · respect any advertising or claims rules binding your sector, and never promise a result you cannot evidence.

12. Forbidden — as prescriptive as the template

Forbidden	Why it kills citability
Dated statistical intro ("according to the 2015 report…")	The engine reads the top first and infers the whole page is of that era
Commercial H2 ("With [BRAND], an expert for all your…")	An H2 that isn't a question answers nothing. There is nothing to extract
CTA before the factual answer	The engine hits conversion copy where the answer should be, and cites whoever answered
Brand-centric opening line, or an acronym used before it is defined	Zero density in the highest-value 250 words; the passage becomes unquotable standalone — and standalone is the only way it gets quoted

13. Delivery checklist (17) + KPIs

text

- [ ]  1. H1 is a question                  - [ ] 10. Every factual claim carries its source
- [ ]  2. Visible "Updated on" under H1     - [ ] 11. No statistic older than 24 months
- [ ]  3. Named author + link to a real bio - [ ] 12. >=5 internal hub-and-spoke links
- [ ]  4. Direct answer in first 250 words  - [ ] 13. "Common mistakes" section present
- [ ]  5. TL;DR bullets present             - [ ] 14. Contextual CTA, not a generic download
- [ ]  6. >=1 real HTML table               - [ ] 15. No promise of a result you can't evidence
- [ ]  7. All H2s phrased as questions      - [ ] 16. Length fits the page's job (see below)
- [ ]  8. >=8 Q/A in the closing FAQ        - [ ] 17. Named author has personally reviewed
- [ ]  9. JSON-LD Article+FAQPage+Person+Org,        and endorsed the copy before publish
          @id-linked

Point 16 depends on the page's job: 2 500-3 500 words on a pillar page (4 000 for a hub) · 1 500-2 200 on a short, urgent-intent product page. On a two-domain setup, the split is prescribed in references/playbooks.md §15 — applying the pillar number to a product page violates it.

KPIs — instrument these; there are no guaranteed targets. Structural metrics are verifiable on day one: visible update date and FAQPage schema on 100% of rewritten pages · mean pillar length. Outcome metrics — citations per engine on your top commercial queries (§10 source 4), AI-Overview citations, organic traffic on rewritten pages — have no defensible benchmark before you have your own: measure the baseline on the first cycle, then set targets against it. A number invented at kickoff is exactly the unsourced claim §11 and point 10 forbid.

14-15. Refresh & multi-domain → references/playbooks.md

5-phase refresh · "What changed" box · slug de-dating + 301 · article↔master-FAQ coupling · two-domain matrix, canonicalisation, anti-cannibalisation.

Troubleshooting

Symptom	Root cause	Fix
IndexNow 403, key file serves fine and matches	Domain not yet validated on the engine console. Local key ≠ declared key	playbooks.md §16 step 4, then re-ping. Sitemap covers crawl meanwhile. Don't touch the script
IndexNow 422 One or more URLs are not related to your verified domain	A <loc> host ≠ INDEXNOW_HOST — usually a staging URL or an unescaped &amp; left in the URL	One host per ping. The script decodes XML entities; check INDEXNOW_HOST matches the sitemap's real host
"I can't access that page" on a URL the user pasted	On-demand fetcher blocked by a blanket *bot* WAF or robots rule	§1 named allowlist — and check the WAF, robots.txt is not the only gate
Rich-results test green, engines still cite a competitor	Several contradictory JSON-LD blocks → no resolvable entity	Collapse to one @graph, @id cross-refs, reused across pages
hreflang group silently ignored	Non-reciprocal alternates, or missing self-reference	§4. Every URL declares all alternates, including itself
Build hangs or fails after adding the ping	An error path throws, or no timeout	All paths resolve(), 10 s timeout, || true on the build command
dateModified recent, engine still treats page as stale	No visible date in the rendered HTML	Render "Updated on …" under the H1
Page rewritten answer-first, still not quoted	The answer isn't self-contained — it leans on the paragraph above	§11 block 3: number + source, readable with no page around it

Scope

This skill ONLY: edits files you own (robots.txt, llms.txt, sitemap, JSON-LD, page copy) · reads public pages you name — your own, and competitors' during §9 · submits your own public sitemap URLs to a public endpoint, opt-in, key from env · produces audit tables and rewrite plans from those public pages · drafts copy for a named human to review and endorse.

This skill NEVER: logs into a console, CMS or third-party account for you · touches paid ranking, paid reviews, or any manipulation of a competitor's listing · reads, scrapes or stores personal data — internal demand signals enter only as an aggregated theme list a human hands you · publishes a claim without a source, or a promise of a result you cannot evidence · attributes drafted copy to a named human who has not reviewed and endorsed it, or fabricates an author, a credential or a bio · names a real competitor in a shareable artefact (A/B/C only) · writes ${INDEXNOW_KEY} to disk outside the key file the protocol requires · advises evading a rate limit, bot detection or any anti-abuse control — if a crawler or console blocks you, stop and involve a human rather than adapting behaviour to get around it. Engine ToS and applicable data law (GDPR or local equivalent) take precedence over every line above.